Deep learning-based dynamic regulation system and method for MICP mine repair process

Abstract

The application discloses a MICP mine repair process dynamic regulation system and method based on deep learning, and particularly relates to the intelligent control technical field of the mine repair process, and is used for solving the problem of uncontrollable repair process and uneven effect of the existing MICP repair technology caused by the dependence on fixed parameters and artificial experience. Visual data and sensing data are obtained in real time by image acquisition equipment and environmental sensors arranged in the repair area; a trained deep learning model is used to identify repair state indicators and extract feature information thereof; the current states of multiple repair targets and their interaction relationship are evaluated based on the feature information to generate an optimized trade-off result; the future process evolution trajectory of the repair target is predicted, a deviation is generated by comparing with an expected trajectory, and a regulation instruction is formed accordingly; and finally, the grouting parameters of the MICP grouting equipment are dynamically adjusted according to the regulation instruction, so that the adaptive closed-loop precise control of the repair process is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology for mine restoration processes, and more specifically, to a deep learning-based MICP dynamic control system and method for mine restoration processes. Background Technology

[0002] Microbial-induced calcium carbonate precipitation (MICP) technology, due to its environmental friendliness, has been explored for applications in engineering fields such as mine slope reinforcement, crack repair, and ecological restoration. Currently, the engineering implementation methods for this technology are primarily based on a defined formula and a pre-set construction plan. Operators pre-set key grouting parameters, such as bacterial concentration and nutrient solution injection rate, based on static reports from geological surveys. During the restoration process, the assessment of the soil and rock conditions (such as crack development, settlement deformation, and biochemical environment) typically relies on periodic manual on-site sampling, followed by laboratory physical measurements and chemical analyses. Data obtained in this way is severely delayed in time and only provides discrete point information, making it difficult to comprehensively and in real-time reflect the complex and dynamically changing real-world conditions within a large restoration area. Therefore, the entire restoration process is essentially an open-loop control process lacking timely and sufficient feedback.

[0003] The fundamental flaw of existing technology lies in placing the dynamic biochemical reaction process of MICP under a static, open-loop control framework. The preset fixed grouting parameters cannot respond to the non-uniform working conditions that change in real time at the repair site, resulting in uncontrollable distribution of the repair agent, reaction process and final reinforcement effect, and making it difficult to guarantee the uniformity and reliability of the repair body. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, this invention provides a deep learning-based dynamic control system and method for the MIP mine restoration process to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A deep learning-based dynamic control method for MICP mine restoration processes includes the following steps:

[0007] S1: Obtain visual and sensor data of the repaired area through image acquisition devices and environmental sensors deployed in the repaired area;

[0008] S2: Input visual data and sensor data into the trained deep learning model to identify and output at least one repair status indicator of the repaired area;

[0009] S3: Extract feature information of the repair status index based on at least one repair status index;

[0010] S4: Based on feature information, evaluate the current state of multiple remediation targets and their interactions, and generate optimization trade-off results;

[0011] S5: Based on feature information, predict the process evolution trajectory of multiple repair targets within a preset time period, transform the optimization trade-off result into the expected co-evolution trajectory and calculate the deviation between the two, and generate control instructions for the MICP grouting equipment based on the deviation.

[0012] S6: Control the MICP grouting equipment to adjust grouting parameters according to the control instructions.

[0013] Furthermore, visual and sensor data of the repaired area are acquired through image acquisition devices and environmental sensors deployed in the repaired area, including:

[0014] Deploy image acquisition equipment to cover the repair area; the image acquisition equipment is used to collect visual data of the repair area.

[0015] Environmental sensors are deployed within the remediation area to collect sensor data from the remediation area.

[0016] Visual data of the repaired area is collected using image acquisition equipment, and sensor data of the repaired area is collected using environmental sensors.

[0017] Furthermore, visual and sensor data are input into a trained deep learning model to identify and output at least one restoration status indicator for the restored region, including:

[0018] Image feature extraction is performed on visual data to obtain image feature data;

[0019] Temporal features are extracted from the sensor data to obtain sensor feature data;

[0020] Image feature data and sensor feature data are fused to form fused feature data;

[0021] Input the fused feature data into the trained deep learning model;

[0022] By processing fused feature data through a trained deep learning model, at least one repair status indicator is identified and output.

[0023] Furthermore, the trained deep learning model is obtained by: acquiring a training dataset containing historical visual data, historical sensor data, and corresponding repair status index labels; using the training dataset to iteratively train the initial deep learning model, adjusting the model parameters through the backpropagation algorithm until the error between the model output and the label is less than a preset error threshold, thus obtaining the trained deep learning model.

[0024] Furthermore, based on at least one repair status indicator, feature information of the repair status indicator is extracted, including:

[0025] Based on at least one repair status index, calculate the rate of change characteristics and spatial distribution characteristics of the repair status index;

[0026] Correlation analysis was performed on the change rate characteristics and spatial distribution characteristics to obtain the dynamic correlation characteristics of the repair status indicators;

[0027] By integrating the characteristics of change rate, spatial distribution, and dynamic correlation, characteristic information of the repair status index is formed.

[0028] Furthermore, based on feature information, the current state of multiple remediation objectives and their interactions are evaluated to generate optimization trade-off results, including:

[0029] Based on feature information, the current state assessment values ​​corresponding to the crack stability repair target, the settlement control repair target, and the ecological restoration repair target are determined respectively;

[0030] Analyze the mutual constraints among the current state assessment values ​​of crack stability repair objectives, settlement control repair objectives, and ecological restoration repair objectives;

[0031] Based on the mutual constraints, the priority weights of the crack stability repair target, the settlement control repair target, and the ecological restoration repair target are dynamically allocated to generate an optimized trade-off result.

[0032] Furthermore, the analysis of the mutual constraints among the current state assessment values ​​of the crack stability repair target, the settlement control repair target, and the ecological restoration repair target includes: inputting the current state assessment values ​​of the crack stability repair target, the settlement control repair target, and the ecological restoration repair target into a multi-objective optimization model; in the multi-objective optimization model, based on the Pareto optimality principle, calculating the degree of influence of the improvement of the state assessment value of any repair target on the state assessment values ​​of the other repair targets, and quantifying the mutual constraints.

[0033] Furthermore, based on feature information, the process evolution trajectory of multiple repair targets within a preset time period is predicted. The optimization trade-off result is transformed into the expected co-evolution trajectory, and the deviation between the two is calculated. Based on the deviation, control instructions for the MICP grouting equipment are generated, including:

[0034] Based on feature information, predict the process evolution trajectory of multiple repair targets within a preset time period in the future;

[0035] Based on the optimization trade-off results, the expected co-evolutionary trajectory of multiple remediation objectives is generated;

[0036] The deviation is calculated by comparing the process evolution trajectory with the expected co-evolution trajectory.

[0037] Based on the deviation, control quantities for adjusting MICP grouting parameters are determined, and control commands for the MICP grouting equipment are generated.

[0038] Furthermore, the grouting parameters of the MICP grouting equipment are adjusted according to the control instructions, including:

[0039] Analyze the control commands to determine the grouting parameters to be adjusted and the corresponding adjustment amounts;

[0040] Based on the adjustment amount, control signals are generated for the bacterial solution supply unit and nutrient solution supply unit in the MICP grouting equipment.

[0041] The bacterial solution supply unit and the nutrient solution supply unit are driven by control signals to adjust the grouting parameters.

[0042] On the other hand, the present invention provides a deep learning-based dynamic control system for the MICP mine restoration process, comprising the following modules:

[0043] The data acquisition module is used to acquire visual and sensor data of the repaired area through image acquisition devices and environmental sensors deployed in the repaired area;

[0044] The indicator output module is used to input visual data and sensor data into the trained deep learning model, identify and output at least one repair status indicator of the repaired area.

[0045] The information extraction module is used to extract feature information of the repair status indicators based on at least one repair status indicator.

[0046] The trade-off generation module is used to evaluate the current state of multiple repair objectives and their interactions based on feature information, and generate optimized trade-off results.

[0047] The instruction generation module is used to predict the process evolution trajectory of multiple repair targets within a preset time period based on feature information, transform the optimization trade-off result into the expected co-evolution trajectory and calculate the deviation between the two, and generate control instructions for the MICP grouting equipment based on the deviation.

[0048] The command control module is used to control the MICP grouting equipment to adjust the grouting parameters according to the control commands.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. By leveraging multi-source sensing data to drive the entire control process, the system achieves dynamic capture and response to the working conditions at the repair site. By acquiring and fusing visual and sensor data in real time, the system can continuously and non-contactly grasp the real state of a large-scale repair area, overcoming the shortcomings of manual sampling lag and insufficient spatial coverage. On this basis, deep learning models are used to intelligently identify and extract features from complex and nonlinear repair states, providing a high-dimensional and accurate information foundation for subsequent decision-making. This transforms the original static and open-loop grouting process into a dynamic process driven by real-time data flow and possessing adaptive capabilities.

[0051] 2. By comprehensively analyzing the real-time status and interrelationships of multiple repair targets, the system can autonomously generate a trade-off strategy that reflects the overall optimality of the project. By predicting the future trajectory of the repair process and comparing it with the expected target, the system can identify deviations in advance and generate forward-looking control instructions. This control strategy based on model prediction and deviation compensation allows for fine-tuning of grouting parameters to address the inhomogeneity and time-varying nature within the repair area. This ensures the spatial uniformity and temporal controllability of the distribution of the repair agent and the biochemical reaction process, thereby significantly improving the reliability of the mechanical properties and long-term stability of the repair. Attached Figure Description

[0052] Figure 1 This is a flowchart of the deep learning-based dynamic control method for the MICP mine restoration process of the present invention.

[0053] Figure 2 This is a schematic diagram of the structure of the MICP mine restoration process dynamic control system based on deep learning, which is based on the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0055] Example 1: Figure 1 This invention presents a deep learning-based dynamic control method for the MIP (Micro-Environmental Protection Process) mine restoration process, comprising the following steps:

[0056] S1: Obtain visual and sensor data of the repaired area through image acquisition devices and environmental sensors deployed in the repaired area;

[0057] S2: Input visual data and sensor data into the trained deep learning model to identify and output at least one repair status indicator of the repaired area;

[0058] S3: Extract feature information of the repair status index based on at least one repair status index;

[0059] S4: Based on feature information, evaluate the current state of multiple remediation targets and their interactions, and generate optimization trade-off results;

[0060] S5: Based on feature information, predict the process evolution trajectory of multiple repair targets within a preset time period, transform the optimization trade-off result into the expected co-evolution trajectory and calculate the deviation between the two, and generate control instructions for the MICP grouting equipment based on the deviation.

[0061] S6: Control the MICP grouting equipment to adjust grouting parameters according to the control instructions.

[0062] S1: Obtain visual and sensor data of the repair area through image acquisition devices and environmental sensors deployed in the repair area. Specifically, this is implemented as follows:

[0063] In the specific implementation process, in order to obtain visual and sensor data of the restoration area, it is necessary to deploy image acquisition equipment and environmental sensors and complete data acquisition. When deploying image acquisition equipment to cover the restoration area, the spatial distribution scheme of the image acquisition equipment should be planned according to the boundary range and terrain features of the restoration area. For example, for a restoration area larger than 1000 square meters, a multi-rotor UAV carrying a visible light camera can be used as the image acquisition equipment. The automatic flight path of the UAV should be planned so that it can vertically cover the restoration area. The spacing of the flight path is determined based on the camera's field of view and the required image overlap rate. For example, a flight path spacing of 15 meters can be set to ensure that adjacent flight path images have a 60% lateral overlap rate, thereby ensuring that the acquired visual data can completely cover the restoration area. When the image acquisition equipment collects visual data of the restoration area, specific operating parameters of the image acquisition equipment should be set. These operating parameters are determined according to the spatial scale requirements of the monitored target. For example, when monitoring a rock crack with a width of millimeters, the ground sampling distance of the image acquisition equipment is set to 2 centimeters. This is achieved by controlling the flight altitude of the UAV and the focal length of the camera. The acquisition time interval is set according to the reaction rate of the restoration process. For example, in the critical reaction stage after the injection of the restoration agent, the acquisition time interval is set to 6 hours.

[0064] When deploying environmental sensors within the remediation area, the type and location of the sensors are determined based on the engineering geological zoning of the area and the spatial differences of the monitoring targets. Within the remediation area, the installation locations of the environmental sensors are determined using a grid-like distribution method or a focused distribution method based on characteristic areas. For example, a set of displacement sensors is symmetrically placed 5 meters apart on both sides of the crack, and pH and temperature sensors are placed at the center and edge of the remediation area, respectively. During installation, the sensor probes are embedded in or in close contact with the soil and rock medium being measured. For example, the glass electrode of the pH sensor is inserted into a pre-drilled pore water sampling hole, and the temperature sensor probe is buried in the shallow soil. When collecting sensor data from the remediation area, the data acquisition frequency and operating mode of the environmental sensors are set. The data acquisition frequency is set according to the expected rate of change of the measured environmental parameter; for example, the acquisition frequency for temperature parameters is set to once every 10 minutes, and the acquisition frequency for pore water pressure parameters is set to once every hour. The numerical range of the sensing data is constrained by the sensor's own range, for example, using a pH sensor with a range of 0 to 14 and a temperature sensor with a range of -20 degrees Celsius to 60 degrees Celsius.

[0065] When acquiring visual data of the restoration area using image acquisition equipment, the equipment is controlled to execute shooting tasks according to a predetermined plan. For example, a drone is controlled to autonomously fly along a predetermined route, hovering at each waypoint and triggering the camera to capture a high-resolution digital image. The camera exposure parameters are automatically adjusted according to the ambient light intensity, and each image is associated with the GPS coordinates and Beijing timestamp of the shooting time. When acquiring sensor data of the restoration area using environmental sensors, the sensors periodically measure according to an internal timer, converting the sensed physical or chemical signals into standard current signals, and then converting them into digital values ​​through an analog-to-digital converter. Each digital value obtained from a measurement, along with the corresponding sensor device number and measurement timestamp, forms a sensor data record. Visual and sensor data are transmitted to a data center via a wireless communication network; for example, visual data is transmitted via the drone's 4G data link, and sensor data is transmitted via a low-power wide-area IoT network. To ensure a unified data time base, all image acquisition equipment and environmental sensors are synchronized with a standard time source before acquisition begins, for example, by calibrating their respective system clocks through a network time protocol server. The acquired raw data is subjected to quality verification, such as checking whether the image is distorted due to motion blur, and whether the sensing data exceeds the effective range of the sensor calibration. For data records that fail verification, the corresponding re-acquisition mechanism is triggered or invalidation is marked.

[0066] S2: Input visual and sensor data into the trained deep learning model to identify and output at least one restoration status indicator for the restored region. Specifically, this is implemented as follows:

[0067] Based on the visual and sensor data of the repair area obtained in the aforementioned steps, the process of identifying and outputting repair status indicators is performed. When extracting image features from the visual data, computer vision methods are used to process the visual data to obtain image feature data. The visual data consists of multiple digital images. First, each digital image is preprocessed, including grayscale conversion and histogram equalization to enhance image contrast. Image feature extraction is specifically implemented using scale-invariant feature transform algorithms or accelerated robust feature algorithms. These algorithms detect key points in the digital images and calculate local feature descriptors for these key points. For example, for visual data reflecting crack morphology, the algorithm identifies pixels with drastic grayscale changes as key points. Then, a neighborhood of a certain pixel size is taken centered on the key point, and the orientation histogram of the pixel gradient within this neighborhood is calculated. This orientation histogram constitutes a 128-dimensional feature vector, which is an instance of the image feature data. Multiple feature vectors extracted from multiple digital images together form the image feature dataset.

[0068] When extracting time-series features from sensor data, the sensor data is processed to obtain sensor feature data. Sensor data is a set of observations in time sequence. First, the sensor data undergoes preprocessing operations such as missing value imputation and noise filtering. Missing value imputation uses linear interpolation, and noise filtering is implemented using a moving average filter. Time-series feature extraction is accomplished by calculating statistics within a sliding time window. The length of the sliding time window is determined based on the physicochemical time constant of the repair process. For example, for pH value sensor data, a sliding time window length of 24 hours is set. Within each sliding time window, the mean, variance, first-order difference mean, and time delay of the first zero point of the autocorrelation function are calculated. These four statistics constitute a 4-dimensional feature vector, which is an instance of the sensor feature data. Multiple feature vectors extracted from multiple environmental sensors under multiple sliding time windows together form the sensor feature dataset.

[0069] When fusing image feature data and sensor feature data, a spatiotemporal alignment relationship is established between the two types of feature data. This spatiotemporal alignment is achieved through a unified timestamp-spatial location mapping table. The table records the acquisition time and geographical range corresponding to each set of image feature data, and the acquisition time and sensor location corresponding to each set of sensor feature data. The fusion process specifically employs a feature-level fusion strategy. For image and sensor feature data that are spatially related within the same time period, their feature vectors are concatenated. For example, a 128-dimensional image feature vector from the same region with a time difference within one hour is concatenated with a 4-dimensional sensor feature vector to form a 132-dimensional fused feature vector. All spatiotemporally aligned fused feature vectors are arranged in chronological order to form a fused feature data sequence.

[0070] When inputting fused feature data into a trained deep learning model, the data needs to be prepared according to the input format specified by the trained deep learning model. The fused feature data sequence is divided into fixed-length samples, each sample containing fused feature vectors for consecutive time steps. For example, each sample contains fused feature vectors for 24 consecutive time steps, with each time step spaced 1 hour apart. Each sample is standardized before being input into the trained deep learning model using the Z-score standardization method, which involves subtracting the mean of the features in the entire training dataset and dividing by the standard deviation. The standardized samples are then input into the trained deep learning model as a tensor data structure.

[0071] A trained deep learning model processes fused feature data to identify and output at least one restoration status indicator. The trained deep learning model receives the input fused feature data tensor and performs calculations through multiple internal nonlinear transformations. The internal structure of the trained deep learning model includes an input layer, multiple hidden layers, and an output layer. The hidden layers consist of fully connected layers and activation functions, and the number of neurons in the output layer equals the number of restoration status indicators to be identified. For example, when identifying three restoration status indicators—crack width, settlement, and pH value—the output layer uses three neurons. The forward propagation calculation process of the trained deep learning model involves the input data starting from the input layer, sequentially passing through the weight matrix multiplication, bias vector addition, and activation function mapping operations of each hidden layer, finally reaching the output layer to produce the original output value. The original output value of the output layer undergoes denormalization, which is the reverse of the previous normalization process, converting the model output value back to a value with actual physical units. These values ​​with actual physical units are the identified restoration status indicators; for example, crack width is output in millimeters, settlement in centimeters, and pH value as a dimensionless number.

[0072] The trained deep learning model is obtained as follows: A training dataset containing historical visual data, historical sensor data, and corresponding repair status indicator labels is acquired. The historical visual and sensor data originate from complete data records collected and archived in the same repair area during historical projects. The corresponding repair status indicator labels are derived from standard values ​​obtained through manual measurement and laboratory analysis within the historical project. The training dataset is constructed by selecting multiple complete time periods from the historical data archive. For each time period, historical visual and sensor data are processed using the same steps as image feature extraction, temporal feature extraction, and feature fusion to generate historical fused feature data samples. Each historical fused feature data sample is paired with the standard value of the repair status indicator for the corresponding time period to form a training sample. The initial deep learning model is iteratively trained using the training dataset. The initial deep learning model uses a multilayer perceptron or one-dimensional convolutional neural network structure, with its network weights and bias parameters randomly initialized. During iterative training, a batch of training samples is input into the initial deep learning model in each iteration. The difference between the model-predicted repair status indicators and the actual repair status indicator labels is calculated, and the difference is quantified using the mean squared error function. Model parameters are adjusted using the backpropagation algorithm, which updates parameter values ​​based on the gradient of the error function with respect to each parameter, combined with an optimizer algorithm. The optimizer algorithm employs stochastic gradient descent or its variants, such as the Adam optimizer, with a learning rate set to 0.001. Iterative training continues until the error between the model output and the label is less than a preset error threshold. This preset error threshold is set according to engineering accuracy requirements, for example, a mean squared error of less than 0.01. Training stops when the training loss no longer decreases significantly over multiple consecutive training epochs. The resulting model with fixed parameters is the trained deep learning model.

[0073] S3: Extract feature information of at least one repair status indicator, specifically as follows:

[0074] Based on at least one repair status index of the repaired area identified and output in the aforementioned steps, a process is performed to extract feature information of the repair status index. When calculating the rate of change and spatial distribution characteristics of the repair status index based on at least one repair status index, the rate of change characteristic is obtained by analyzing how quickly the repair status index changes over time, and the spatial distribution characteristic is obtained by analyzing the spatial differences of the repair status index within the repaired area.

[0075] The specific process for calculating the rate of change characteristics is as follows: Each repair status index is treated as a time series. For a time series of a repair status index, the index values ​​at two consecutive time points are taken. The difference between the index value at the latter time point and the value at the former time point is calculated. This difference is then divided by the time interval between the two time points. The quotient obtained is the instantaneous rate of change of the repair status index within that time interval. For example, for the crack width index, the value at time point T1 is 10 mm, and the value at time point T2 is 10.5 mm. The interval between T1 and T2 is 6 hours. The instantaneous rate of change is 0.5 mm divided by 6 hours, which equals 0.0833 mm per hour. This process is repeated throughout the entire time series, calculating the instantaneous rate of change between all adjacent time points to form a rate of change time series. Statistical features are extracted from the rate of change time series as rate of change characteristics. These statistical features include the mean, standard deviation, maximum value, and minimum value of the rate of change time series. For example, a time series of the rate of change of a crack width index over 24 hours might have a mean of 0.05 mm / hour, a standard deviation of 0.02 mm / hour, a maximum of 0.1 mm / hour, and a minimum of 0.01 mm / hour. These four statistics together constitute an instance of the rate of change characteristic of the crack width index. If multiple repair status indicators exist, the above calculation is performed independently for each indicator to obtain the rate of change characteristic corresponding to each indicator.

[0076] The specific process for calculating spatial distribution characteristics involves basing the calculation on the spatial location information corresponding to each restoration status index value. Each restoration status index value is associated with a spatial coordinate at the time of generation; this coordinate comes from the deployment location of the image acquisition device or environmental sensor that collected the data source. All restoration status index values ​​with the same spatial coordinates are marked on a plan view of the restoration area. Spatial distribution characteristics are obtained by calculating the degree of variability and gradient characteristics of this type of restoration status index value in space. The degree of variability is characterized using the coefficient of variation in spatial statistics; that is, the standard deviation of the restoration status index value at all spatial points is calculated first, and then divided by the average value of the index value at all spatial points; the resulting ratio is the spatial coefficient of variation. Gradient characteristics are characterized by calculating the spatial gradient; that is, multiple directions are selected within the restoration area, and the change in restoration status index value per unit distance in each direction is calculated. For example, for the pH value index, an east-west axis is set within the restoration area, and the change in pH value per meter traveled from the western boundary to the eastern boundary is calculated; this change is the east-west spatial gradient of the pH value. The spatial distribution characteristics of a remediation status indicator are specifically composed of three values: the spatial coefficient of variation, the east-west spatial gradient, and the north-south spatial gradient. For example, the spatial distribution characteristics of pH value might be: a spatial coefficient of variation of 0.15, an east-west spatial gradient of 0.05 per meter, and a north-south spatial gradient of 0.03 per meter. The above spatial distribution calculation is performed independently for each remediation status indicator to be analyzed, yielding the spatial distribution characteristics corresponding to each indicator.

[0077] When performing correlation analysis on the rate of change characteristics and spatial distribution characteristics to obtain the dynamic correlation characteristics of repair status indicators, the correlation analysis aims to reveal the intrinsic relationship between the rate of change of repair status indicators over time and the non-uniformity of their spatial distribution. Dynamic correlation characteristics are obtained by calculating the correlation measure between each pair of rate of change characteristics and spatial distribution characteristics. Specifically, for a given repair status indicator, a data pair is formed by combining one component of its rate of change characteristic with one component of its spatial distribution characteristic, and the correlation coefficient of this data pair over multiple consecutive time windows is calculated. The correlation coefficient is calculated using the Pearson product-moment correlation coefficient formula, which calculates the covariance of the two data sequences and then divides it by the product of the standard deviations of the two data sequences. For example, analyzing the correlation between the mean rate of change of crack width and the spatial coefficient of variation of crack width. First, the mean rate of change of crack width over the past 5 time windows (6 hours each) is obtained, denoted as sequence A. Simultaneously, the spatial coefficient of variation of crack width for the corresponding 5 time windows is obtained, denoted as sequence B. Sequence A and sequence B each contain 5 values. Calculate the Pearson product-moment correlation coefficient between sequence A and sequence B. If the result is 0.8, it indicates a strong positive correlation between the mean rate of change of crack width and its spatial coefficient of variation; that is, the faster the change, the greater the spatial difference. This correlation coefficient value of 0.8 constitutes a dynamic correlation feature. Iterate through all repair status indicators and all possible meaningful combinations between each component of the rate of change feature and each component of the spatial distribution feature for each indicator, calculating their correlation coefficients. The resulting set of correlation coefficients is the dynamic correlation feature of the repair status indicators. The dynamic correlation feature is ultimately represented as a vector containing multiple correlation coefficient values.

[0078] When integrating the rate of change features, spatial distribution features, and dynamic correlation features to form the feature information of the repair status index, a vector concatenation method is used to merge the three types of features into a single high-dimensional feature vector. First, the rate of change feature of each repair status index is represented as a sub-vector; for example, the crack width rate of change feature sub-vector is [mean, standard deviation, maximum value, minimum value]. All the rate of change feature sub-vectors of the repair status indexes are concatenated according to a predetermined index order to form the total rate of change feature vector. Second, the spatial distribution feature of each repair status index is represented as a sub-vector; for example, the crack width spatial distribution feature sub-vector is [spatial coefficient of variation, east-west gradient, north-south gradient]. All the spatial distribution feature sub-vectors of the repair status indexes are concatenated according to the same predetermined index order to form the total spatial distribution feature vector. Next, the dynamic correlation feature, which is already a correlation coefficient vector, is directly used as the dynamic correlation feature vector. Finally, the total rate of change feature vector, the total spatial distribution feature vector, and the dynamic correlation feature vector are concatenated end-to-end in terms of dimension. For example, suppose there are two repair status indicators, with an 8-dimensional feature vector representing the rate of change, a 6-dimensional feature vector representing the spatial distribution, and a 10-dimensional feature vector representing the dynamic correlation. The concatenated feature vector of these indicators forms a 24-dimensional feature vector. This final feature vector fully encodes the state information of the repair status indicator across three dimensions: time, space, and spatiotemporal coupling, and is used as input to subsequent decision-making processes. Each dimension of the feature vector has a clear physical or statistical meaning, corresponding to a specific feature metric.

[0079] S4: Based on feature information, evaluate the current state of multiple remediation targets and their interactions, and generate optimization trade-off results. The specific implementation is as follows:

[0080] Based on the feature information of the restoration status indicators extracted in the aforementioned steps, the process of evaluating the current state of multiple restoration targets and their interactions, and generating optimized trade-off results, is performed. When determining the current state assessment values ​​corresponding to the crack stability restoration target, the settlement control restoration target, and the ecological restoration target based on the feature information, it is necessary to extract the feature components directly related to each restoration target from the feature information and perform comprehensive calculations. For the crack stability restoration target, its current state assessment value is obtained by calculating the weighted sum of the crack-related feature components in the feature information. Specifically, all components representing the crack state in the feature information vector are identified, such as the mean crack width change rate and the crack width spatial variation coefficient. An importance weight is assigned to each identified crack-related feature component. The importance weight is pre-set based on engineering experience or expert scoring methods; for example, the importance weight of the mean crack width change rate component is set to 0.6, and the importance weight of the crack width spatial variation coefficient component is set to 0.4. Then, the value of each crack-related feature component is multiplied by its corresponding importance weight, and all products are summed to obtain the current state assessment value of the crack stability restoration target. The current state assessment value is normalized to a range of 0 to 1. The closer the value is to 1, the worse the crack stability; the closer it is to 0, the better the stability. The normalization process is achieved through linear scaling, which involves first calculating the maximum and minimum values ​​of the weighted sum in historical data, then subtracting the minimum value from the current weighted sum, and finally dividing by the difference between the maximum and minimum values.

[0081] For settlement control and remediation targets, the current state assessment value is obtained by calculating the weighted sum of settlement-related feature components in the feature information. All components representing settlement state in the feature information vector are identified, such as the mean rate of settlement change and the spatial gradient of settlement. An importance weight is assigned to each settlement-related feature component; for example, the importance weight of the mean rate of settlement change is set to 0.5, the importance weight of the east-west spatial gradient of settlement is set to 0.3, and the importance weight of the north-south spatial gradient of settlement is set to 0.2. The value of each settlement-related feature component is multiplied by its corresponding importance weight, and all products are summed to obtain the current state assessment value of the settlement control and remediation target. This current state assessment value is also normalized to the range of 0 to 1; the closer the value is to 1, the more severe the settlement control situation; the closer it is to 0, the better the control effect. The normalization method is the same as that used for crack stability remediation targets, but the historical maximum and minimum values ​​of the settlement-related feature components are used for calculation.

[0082] For ecological restoration and repair targets, the current state assessment value is obtained by calculating the weighted sum of ecologically relevant feature components in the feature information. All components representing ecological state in the feature information vector are identified, such as the mean rate of pH change and the spatial coefficient of pH variation. An importance weight is assigned to each ecologically relevant feature component; for example, the importance weight of the mean rate of pH change is set to 0.7, and the importance weight of the spatial coefficient of pH variation is set to 0.3. The value of each ecologically relevant feature component is multiplied by its corresponding importance weight, and all products are summed to obtain the current state assessment value of the ecological restoration and repair target. This current state assessment value is also normalized to the range of 0 to 1; the closer the value is to 1, the further the ecological condition deviates from the ideal state, and the closer it is to 0, the better the ecological condition. The normalization method is the same as for the aforementioned targets, but the historical statistical range of the ecologically relevant feature components is used.

[0083] When analyzing the interrelationships among the current state assessment values ​​of the crack stability restoration, settlement control, and ecological restoration objectives, a multi-objective optimization theory and method are employed for quantitative analysis. Specifically, the current state assessment values ​​of these objectives are input into a multi-objective optimization model. This model aims to analyze the impact of improving the current state assessment value of one restoration objective on the current state assessment values ​​of the others. Based on the Pareto optimality principle, the multi-objective optimization model explores the interrelationships by constructing a virtual optimization process. In this virtual optimization process, an objective function is defined that attempts to minimize the current state assessment value of the third restoration objective while keeping the current state assessment values ​​of the other two objectives unchanged or allowing for small fluctuations. Then, the minimum compromise amount required for the current state assessment values ​​of the other two restoration objectives to improve the current state assessment value of the third objective by a specific small unit, such as a reduction of 0.01, is calculated. This minimum compromise amount quantifies the degree of influence of the third restoration objective on the other two objectives. For example, to reduce the current state assessment value of the crack stability restoration target from 0.5 to 0.49 by 0.01, calculations show that the current state assessment value of the settlement control restoration target needs to increase from at least 0.3 to 0.305, and the current state assessment value of the ecological restoration target needs to increase from at least 0.4 to 0.402. Therefore, the impact of the crack stability restoration target on the settlement control restoration target is quantified as 0.005 / 0.01 = 0.5, and the impact on the ecological restoration target is quantified as 0.002 / 0.01 = 0.2. This process is repeated to calculate the impact of improving the current state assessment value of the settlement control restoration target on the crack stability restoration target and the ecological restoration target, and the impact of improving the current state assessment value of the ecological restoration target on the other two targets. Finally, a set of six impact coefficients is obtained, which completely quantify the pairwise constraints between the three restoration targets, forming a constraint matrix.

[0084] Based on the mutual constraints, the priority weights of the crack stability repair objective, the settlement control repair objective, and the ecological restoration repair objective are dynamically allocated to generate an optimized trade-off result. The dynamic allocation process is based on the constraint matrix calculated at the current moment and the current state evaluation value of each repair objective. The goal of priority weight allocation is to find a weight combination that, under limited resources, guides subsequent regulatory actions to most effectively coordinate the three conflicting objectives. The specific allocation algorithm adopts a decision-making method based on distance from an ideal point. First, an ideal point is constructed, whose coordinates are composed of the best evaluation values ​​that each of the three repair objectives can independently achieve at the current engineering stage. These best evaluation values ​​are derived from historical successful cases or theoretical calculations. For example, at the current repair stage, the possible best evaluation value for the crack stability repair objective is 0.2, the possible best evaluation value for the settlement control repair objective is 0.25, and the possible best evaluation value for the ecological restoration repair objective is 0.3. Second, the current state evaluation value of the three repair objectives is regarded as an actual point in three-dimensional space. The Euclidean distance from the actual point to the ideal point is calculated. The principle of dynamic weight allocation is to make the weight allocation tend to reduce the distance between the actual point and the ideal point. The weights are inversely proportional to the influence coefficients in the constraint matrix and directly proportional to the degree to which the current state assessment deviates from its ideal value. That is, if the influence coefficient of a remediation target is large when it improves on other targets, it indicates a high cost to improve it, and a lower priority weight is temporarily assigned; if the current state assessment of a remediation target deviates significantly from its ideal value, it indicates a more urgent problem, and a higher priority weight is assigned. The specific weight calculation formula is as follows: the initial weight of each remediation target is equal to the absolute value of the difference between its current state assessment value and its ideal value, divided by the sum of the absolute values ​​of such differences for the three targets, to obtain a basic weight. Then, this basic weight is corrected using the average influence coefficient of that target on other targets in the constraint matrix, by dividing the basic weight by 1 and adding the average influence coefficient. Finally, the three corrected weights are normalized so that their sum equals 1. For example, the calculated dynamic weight for the crack stability remediation target is 0.5, the dynamic weight for the settlement control remediation target is 0.3, and the dynamic weight for the ecological restoration remediation target is 0.2. These three weight values ​​together constitute the optimization trade-off result, which clearly indicates which repair objective and to what extent the resources should be allocated to it under the current state. The optimization trade-off result is a vector containing the three weight values, which will be passed as a key input to subsequent decision-making steps.

[0085] S5: Based on feature information, predict the process evolution trajectory of multiple repair targets within a preset time period, transform the optimization trade-off result into the expected co-evolution trajectory and calculate the deviation between the two, and generate control instructions for the MICP grouting equipment based on the deviation. The specific implementation is as follows:

[0086] Based on the feature information and optimization trade-offs of the repair status indicators obtained in the aforementioned steps, the process of predicting the progress of repair targets, generating expected trajectories, calculating deviations, and generating control commands is executed. When predicting the evolution trajectories of multiple repair targets within a future preset time period based on feature information, a time series prediction method is used to evaluate the future state of the repair targets. Specifically, the feature information is input into a trained time series prediction model, which outputs a sequence of predicted state evaluation values ​​for multiple repair targets at a series of equally spaced time points within the future preset time period. The length of the future preset time period is set according to the phased goals and control cycle of the repair project; for example, the future preset time period is set to 24 hours. The training process of the time series prediction model involves collecting feature information data from different time points in historical projects and corresponding future time period actual repair target state evaluation value data to form training sample pairs. Using these training sample pairs, the internal parameters of the time series prediction model are trained by minimizing the error between the predicted and actual values ​​until the model prediction error is less than a preset prediction error threshold. During prediction, the current moment's feature information, i.e., a multi-dimensional feature vector, is input into the trained time series prediction model. The model sequentially outputs the predicted state assessment values ​​for the crack stability repair target, settlement control repair target, and ecological restoration target from the 1st hour, 2nd hour, up to the 24th hour after the current moment. These three predicted state assessment values ​​for each future time point constitute a three-dimensional coordinate point. Connecting all the three-dimensional coordinate points of all future time points in chronological order forms the process evolution trajectories of the crack stability repair target, settlement control repair target, and ecological restoration target, respectively. Each process evolution trajectory is a discrete time series, representing the possible development path of each repair target within the next 24 hours if no intervention is made in the current grouting process.

[0087] When generating the expected co-evolutionary trajectory of multiple remediation objectives based on the optimization trade-off results, the dynamically allocated priority weights in the optimization trade-off results are transformed into specific expectations of the improvement rate of each remediation objective's state. The optimization trade-off result is a vector containing three weight values; for example, the weight of the crack stability remediation objective is 0.5, the weight of the settlement control remediation objective is 0.3, and the weight of the ecological restoration remediation objective is 0.2. The principle for generating the expected co-evolutionary trajectory is that the higher the weight of the remediation objective, the faster its expected rate of convergence to the ideal value should be. First, the ideal state evaluation values ​​of the three remediation objectives are determined. These ideal values ​​are derived from engineering design requirements or historical best practice data. Then, the difference between the current state evaluation value and the ideal state evaluation value of each remediation objective is calculated. The expected co-evolutionary trajectory is defined as a path that starts from the current state evaluation value and approaches the ideal state evaluation value at a specific rate within a preset future time period. This approach rate is proportional to the priority weight assigned to the remediation objective. Specifically, for each remediation objective, its expected improvement rate is set to be equal to the priority weight of that objective multiplied by a baseline improvement rate constant. The baseline improvement rate constant is determined comprehensively based on the maximum theoretical effectiveness of the repair technology and engineering safety constraints, for example, set to improve by 0.3 state assessment units every 24 hours. Therefore, the expected improvement rate for crack stability repair is 0.5 × 0.3 = 0.15 every 24 hours. Next, starting from the current state assessment value, the expected state assessment value at each time point within a preset future period is linearly extrapolated according to the calculated expected improvement rate. For example, if the current state assessment value for crack stability repair is 0.5 and the expected improvement rate is 0.15 every 24 hours, then the expected state assessment value at the 12th hour should be 0.5 - 0.15 × (12 / 24) = 0.425. Repeating the calculation for all time points and all repair targets yields three expected co-evolution trajectories, which reflect the ideal and coordinated development direction of the three repair targets under the guidance of optimization trade-offs.

[0088] The deviation is calculated by comparing the process evolution trajectory with the expected co-evolution trajectory. The deviation is defined as the difference between the predicted state assessment value and the expected state assessment value at the same future time point. The calculation process is performed independently for each repair target. For a given repair target, such as crack stability repair, a sequence of predicted state assessment values ​​at N time points within a preset future time period is extracted from its process evolution trajectory, denoted as sequence P. Simultaneously, a sequence of expected state assessment values ​​at the same N time points on its expected co-evolution trajectory is extracted, denoted as sequence D. The deviation calculation is not a simple single-point difference but considers the comprehensive difference in the entire trajectory morphology. One implementation method is to calculate the root mean square error between the two sequences as the total deviation. Specifically, for each pair of corresponding time point values ​​in sequences P and D, their difference is calculated, then the squares of all differences are summed, divided by the number of time points N, and finally the square root is taken to obtain the total deviation value for the repair target. Furthermore, to guide regulation, the temporal distribution characteristics of the deviation also need to be calculated, i.e., identifying future time intervals where the deviation increases significantly. For example, the instantaneous difference between sequences P and D at each time point is calculated to form an instantaneous deviation sequence. By analyzing the instantaneous deviation sequence, we can pinpoint the time period in the future where the deviation between the predicted and expected states is most severe. Ultimately, the calculated deviation output is a data structure containing the total deviation value for each remediation target and the instantaneous deviation sequence for each target. This deviation data structure quantifies the degree and pattern to which the actual process will deviate from the expected coordinated path if no intervention measures are taken.

[0089] Based on the deviation, when determining the control quantities used to adjust the MICP grouting parameters and generating control commands for the MICP grouting equipment, the process of determining the control quantities is a process of mapping deviation information to specific execution parameters. The core of the control commands is adjusting two key grouting parameters: bacterial solution concentration and nutrient solution injection rate. First, a mapping relationship model between the deviation and the control quantities is established. This mapping relationship model is constructed based on the understanding of the mechanistic mechanism of the MIP remediation process or historical control experience data. The mapping relationship model specifies how deviations of different magnitudes and patterns should be converted into adjustments to bacterial solution concentration and nutrient solution injection rate. Specifically, a control response coefficient is set for the total deviation value of each remediation target. For example, the total deviation control response coefficient for the crack stability remediation target is A1, for the settlement control remediation target is A2, and for the ecological restoration remediation target is A3. These coefficients represent the intensity of grouting parameter adjustment required per unit of total deviation, and their values ​​are determined through experimental calibration or process simulation. Then, the baseline adjustment for bacterial concentration is calculated. This baseline adjustment equals the weighted sum of the total deviations of the three remediation targets multiplied by their corresponding control response coefficients. The weights represent the priority weights in the optimization trade-offs. Simultaneously, the instantaneous deviation sequence is analyzed. If the instantaneous deviation consistently exceeds a preset instantaneous deviation threshold within a future time interval, an additional compensation adjustment is triggered. This compensation adjustment aims to proactively address any impending significant deviations. Ultimately, the total adjustment for bacterial concentration equals the sum of the baseline adjustment and the compensation adjustment. The adjustment for the nutrient solution injection rate is determined proportionally to the bacterial concentration adjustment to maintain a pre-defined stoichiometric ratio between the two, such as a 1:1 molar ratio of urea to calcium ions. After determining the control values, specific control instructions are generated. A control instruction is a data structure or signal containing clearly defined operational parameters. For example, the instruction might be: increase the output concentration of the bacterial solution supply unit by X mg / L, increase the injection rate of the nutrient solution supply unit by Y L / min, and set the instruction validity period to the next Z hours. Here, X, Y, and Z represent the calculated control values ​​and their durations of action. The control command is encapsulated in a standard communication protocol and is ready to be sent to the control system of the MICP grouting equipment for execution.

[0090] The prediction error threshold is set based on the evaluation of model performance during historical training. During model training, a root mean square error sequence between model predictions and true values ​​is calculated using an independent validation dataset. By analyzing the statistical distribution of this error sequence (e.g., taking the 95th percentile) and considering the maximum acceptable error range for state prediction in the remediation project, a specific value is determined as the prediction error threshold. This threshold is used to determine whether the model has been trained sufficiently; training stops when the training error falls below this threshold.

[0091] The instantaneous deviation threshold is set based on a safety margin for fluctuations in the target state. By analyzing a large amount of historical engineering data, the normal fluctuation range of the target state assessment value is calculated, and its standard deviation is used as a benchmark. Simultaneously, considering the control sensitivity of the repair process and the cost of adjustment actions, the threshold is set to 1.5 to 2 times the standard deviation of the normal fluctuation range. This threshold defines the allowable deviation boundary without emergency intervention. When the instantaneous deviation between the predicted trajectory and the expected trajectory exceeds this boundary, a compensation control mechanism is triggered.

[0092] S6: Adjust the grouting parameters of the MICP grouting equipment according to the control instructions. Specifically, the implementation is as follows:

[0093] Based on the control instructions generated in the preceding steps, the final control of the MICP grouting equipment is executed to adjust the grouting parameters. When parsing the control instructions to determine the grouting parameters to be adjusted and their corresponding adjustment amounts, the control instructions are structured data messages. These data messages contain three parts: a header, a body, and a checksum. The header defines the instruction type and version; for example, the instruction type is identified as parameter adjustment, and the version number is 1. The body is the core data area, explicitly listing the names of the grouting parameters to be adjusted and their target adjustment amounts in key-value pairs. The grouting parameters to be adjusted include at least the bacterial concentration and the nutrient solution injection rate. For example, the body might contain one key-value pair for bacterial concentration adjustment: +50 mg / L, and another for nutrient solution injection rate adjustment: +0.5 L / min. The parsing process first verifies the checksum to ensure no errors occurred during transmission. The checksum is calculated using a cyclic redundancy check algorithm. After successful verification, the parsing program reads the body and, based on a predefined parameter name dictionary, identifies and extracts the values ​​corresponding to the two keys: bacterial concentration adjustment and nutrient solution injection rate adjustment. These values ​​represent the adjustment amounts corresponding to the grouting parameters to be adjusted. An adjustment amount is a physical quantity with a positive or negative sign and a specific numerical value; a positive sign indicates an increase, and a negative sign indicates a decrease. The analysis process also checks whether the adjustment amount is within a pre-set safe operating range. For example, the absolute value of the bacterial solution concentration adjustment amount must not exceed 200 mg / L, and the absolute value of the nutrient solution injection rate adjustment amount must not exceed 2 L / min. If the adjustment amount exceeds the safe range, an error handling procedure is triggered; otherwise, the confirmed adjustment amount is passed to the next step.

[0094] When generating control signals for the bacterial solution supply unit and nutrient solution supply unit in the MICP grouting equipment based on the adjustment amount, the control signals are electrical signals or digital instructions that can be directly recognized and executed by industrial actuators. The core of generating control signals is converting the adjustment amount, expressed in engineering units, into specific control quantities that drive the actuator's movement. For the bacterial solution supply unit, its core actuator might be a metering pump or a regulating valve. The control signal needs to specify the target position or target operating speed of this actuator. First, a conversion relationship between the adjustment amount and the actuator's control quantity needs to be established. This relationship is obtained through equipment calibration; for example, by experimentally determining that for every 1 mg / L increase in bacterial solution concentration, the metering pump speed needs to be increased by X revolutions per minute, or the regulating valve opening needs to be increased by Y percentage. Assuming the calibration relationship is linear, the process of generating the control signal for the bacterial solution supply unit is to multiply the analytically obtained bacterial solution concentration adjustment amount by the conversion coefficient K1 to obtain a preliminary control quantity change value. Then, the current operating status of the bacterial solution supply unit is queried to obtain its current control quantity, such as the current metering pump speed. The current control quantity is added to the initial control quantity change value to obtain the target control quantity. Finally, this target control quantity is encapsulated into a specific control signal according to the communication protocol format required by the bacterial culture supply unit controller. For example, if the controller supports 4-20mA analog current signal control, the target control quantity needs to be proportionally converted into a current value within the 4-20mA range; this current value is the final control signal. If the controller supports digital communication protocols such as Modbus, a request frame containing the target control quantity register address and data value needs to be generated. For the nutrient solution supply unit, the generation process of its control signal is completely similar, but different conversion coefficient K2 and current status query are used. In addition, it is necessary to ensure the synchronization of bacterial culture and nutrient solution adjustments; therefore, the two generated control signals will be associated with the same timestamp and prepared to be sent simultaneously or in a predetermined order.

[0095] When using control signals to drive the bacterial solution supply unit and nutrient solution supply unit to adjust grouting parameters, the driving process means actually sending the generated control signals to the equipment and making them effective. This process is completed through an industrial control network. First, the control signal is sent to the corresponding equipment controller. For example, the control signal of the bacterial solution supply unit is converted into a 4-20mA current signal through an analog output board. This current signal is directly connected to the speed setpoint terminal of the frequency converter of the bacterial solution metering pump via a cable. When the current signal changes from the original 12mA to a new 14mA, the frequency converter senses the input change and adjusts the output frequency, thereby changing the speed of the metering pump motor and adjusting the pumping flow rate. The direct result of the flow rate adjustment is a change in the concentration of bacterial solution delivered to the grouting pipeline. For nutrient solution supply units that support digital communication, the control signal is sent as a Modbus write register command to the intelligent controller of the nutrient solution flow regulating valve via an RS-485 bus. After receiving the command, the controller parses the target opening value and drives its internal motor to rotate the valve from the current opening to the target opening, thereby changing the flow rate in the nutrient solution pipeline. To ensure the accuracy and reliability of the adjustments, the drive process typically includes a closed-loop verification step. After the control signal is issued, following a brief stabilization delay, such as 10 seconds, the system rereads the actual feedback values ​​from the bacterial solution supply unit and the nutrient solution supply unit. For example, it reads the actual injection rate via a flow meter and the actual bacterial solution concentration via an online concentration meter. These actual feedback values ​​are compared with the expected adjusted target values. If the deviation exceeds an allowable execution error threshold, such as a difference of more than 5 mg / L between the actual bacterial solution concentration and the target value, a fine-tuning compensation process is triggered. This generates a small compensation control signal for secondary drive until the actual parameters meet the requirements. Finally, when both injection parameters—bacterial solution concentration and nutrient solution injection rate—stabilize at the new operating point required by the control command, the entire process of adjusting the injection parameters is complete. The adjusted injection parameters will act on the repair area, thereby affecting the progress of the MICP response, forming a complete closed loop from perception and decision-making to execution.

[0096] The execution error threshold is set based on the equipment control precision and the allowable fluctuation range of the process. First, control deviation distribution data for the bacterial solution supply unit and nutrient solution supply unit under steady-state conditions are obtained through multiple calibration experiments, and their standard deviation is calculated. Then, considering the minimum requirements for the stability of grouting parameters in the remediation process, such as the bacterial solution concentration fluctuation not affecting the calcium carbonate precipitation rate by more than 5%, a maximum allowable deviation is determined. The execution error threshold is typically set to the smaller of 3 to 4 times the standard deviation of the calibration experiments and the maximum allowable deviation of the process, ensuring reliable control without causing oscillations due to excessive pursuit of precision.

[0097] Example 2: Figure 2A schematic diagram of the structure of the deep learning-based dynamic control system for MICP mine restoration process of the present invention is given. The deep learning-based dynamic control system for MICP mine restoration process includes the following modules:

[0098] The data acquisition module is used to acquire visual and sensor data of the repaired area through image acquisition devices and environmental sensors deployed in the repaired area;

[0099] The indicator output module is used to input visual data and sensor data into the trained deep learning model, identify and output at least one repair status indicator of the repaired area.

[0100] The information extraction module is used to extract feature information of the repair status indicators based on at least one repair status indicator.

[0101] The trade-off generation module is used to evaluate the current state of multiple repair objectives and their interactions based on feature information, and generate optimized trade-off results.

[0102] The instruction generation module is used to predict the process evolution trajectory of multiple repair targets within a preset time period based on feature information, transform the optimization trade-off result into the expected co-evolution trajectory and calculate the deviation between the two, and generate control instructions for the MICP grouting equipment based on the deviation.

[0103] The command control module is used to control the MICP grouting equipment to adjust the grouting parameters according to the control commands.

[0104] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0105] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0106] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0107] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0108] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0109] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0110] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based dynamic control method for MICP mine restoration process, characterized in that, Includes the following steps: S1: Obtain visual and sensor data of the repaired area through image acquisition devices and environmental sensors deployed in the repaired area; S2: Input visual data and sensor data into the trained deep learning model to identify and output at least one repair status indicator of the repaired area; S3: Extract feature information of the repair status index based on at least one repair status index; S4: Based on feature information, evaluate the current state of multiple remediation targets and their interactions, and generate optimization trade-off results; S5: Based on feature information, predict the process evolution trajectory of multiple repair targets within a preset time period, transform the optimization trade-off result into the expected co-evolution trajectory and calculate the deviation between the two, and generate control instructions for the MICP grouting equipment based on the deviation. S6: Control the MICP grouting equipment to adjust grouting parameters according to the control instructions.

2. The method for dynamic control of the MICP mine restoration process based on deep learning according to claim 1, characterized in that, Visual and sensor data of the repaired area are acquired through image acquisition devices and environmental sensors deployed in the repaired area, including: Deploy image acquisition equipment to cover the repair area; the image acquisition equipment is used to collect visual data of the repair area. Environmental sensors are deployed within the remediation area to collect sensor data from the remediation area. Visual data of the repaired area is collected using image acquisition equipment, and sensor data of the repaired area is collected using environmental sensors.

3. The method for dynamic control of the MICP mine restoration process based on deep learning according to claim 1, characterized in that, Visual and sensor data are input into a trained deep learning model to identify and output at least one restoration status indicator for the restored region, including: Image feature extraction is performed on visual data to obtain image feature data; Temporal features are extracted from the sensor data to obtain sensor feature data; Image feature data and sensor feature data are fused to form fused feature data; Input the fused feature data into the trained deep learning model; By processing fused feature data through a trained deep learning model, at least one repair status indicator is identified and output.

4. The deep learning-based dynamic control method for MIP mine restoration process according to claim 3, characterized in that, The trained deep learning model is obtained by: acquiring a training dataset containing historical visual data, historical sensor data, and corresponding repair status index labels; using the training dataset to iteratively train the initial deep learning model, adjusting the model parameters through the backpropagation algorithm until the error between the model output and the label is less than a preset error threshold, thus obtaining the trained deep learning model.

5. The deep learning-based dynamic control method for MIP mine restoration process according to claim 1, characterized in that, Based on at least one repair status indicator, extract feature information of the repair status indicator, including: Based on at least one repair status index, calculate the rate of change characteristics and spatial distribution characteristics of the repair status index; Correlation analysis was performed on the change rate characteristics and spatial distribution characteristics to obtain the dynamic correlation characteristics of the repair status indicators; By integrating the characteristics of change rate, spatial distribution, and dynamic correlation, characteristic information of the repair status index is formed.

6. The method for dynamic control of the MICP mine restoration process based on deep learning according to claim 1, characterized in that, Based on feature information, the current state of multiple remediation objectives and their interactions are evaluated, generating optimization trade-offs, including: Based on feature information, the current state assessment values ​​corresponding to the crack stability repair target, the settlement control repair target, and the ecological restoration repair target are determined respectively; Analyze the mutual constraints among the current state assessment values ​​of crack stability repair objectives, settlement control repair objectives, and ecological restoration repair objectives; Based on the mutual constraints, the priority weights of the crack stability repair target, the settlement control repair target, and the ecological restoration repair target are dynamically allocated to generate an optimized trade-off result.

7. The deep learning-based dynamic control method for MICP mine restoration process according to claim 6, characterized in that, The analysis of the mutual constraints among the current state assessment values ​​of the crack stability repair target, the settlement control repair target, and the ecological restoration repair target includes: inputting the current state assessment values ​​of the crack stability repair target, the settlement control repair target, and the ecological restoration repair target into a multi-objective optimization model; in the multi-objective optimization model, based on the Pareto optimality principle, calculating the degree of influence of the improvement of the state assessment value of any repair target on the state assessment values ​​of the other repair targets, and quantifying the mutual constraints.

8. The method for dynamic control of the MICP mine restoration process based on deep learning according to claim 1, characterized in that, Based on feature information, the process evolution trajectory of multiple repair targets within a preset time period is predicted. The optimization trade-off result is transformed into the expected co-evolution trajectory, and the deviation between the two is calculated. Based on the deviation, control instructions for the MICP grouting equipment are generated, including: Based on feature information, predict the process evolution trajectory of multiple repair targets within a preset time period in the future; Based on the optimization trade-off results, the expected co-evolutionary trajectory of multiple remediation objectives is generated; The deviation is calculated by comparing the process evolution trajectory with the expected co-evolution trajectory. Based on the deviation, control quantities for adjusting MICP grouting parameters are determined, and control commands for the MICP grouting equipment are generated.

9. The method for dynamic control of MICP mine restoration process based on deep learning according to claim 1, characterized in that, The MICP grouting equipment is controlled according to the control instructions to adjust the grouting parameters, including: Analyze the control commands to determine the grouting parameters to be adjusted and the corresponding adjustment amounts; Based on the adjustment amount, control signals are generated for the bacterial solution supply unit and nutrient solution supply unit in the MICP grouting equipment. The bacterial solution supply unit and the nutrient solution supply unit are driven by control signals to adjust the grouting parameters.

10. A deep learning-based dynamic control system for MICP mine restoration process, used to implement the deep learning-based dynamic control method for MICP mine restoration process as described in any one of claims 1-9, characterized in that, Includes the following modules: The data acquisition module is used to acquire visual and sensor data of the repaired area through image acquisition devices and environmental sensors deployed in the repaired area; The indicator output module is used to input visual data and sensor data into the trained deep learning model, identify and output at least one repair status indicator of the repaired area. The information extraction module is used to extract feature information of the repair status indicators based on at least one repair status indicator. The trade-off generation module is used to evaluate the current state of multiple repair objectives and their interactions based on feature information, and generate optimized trade-off results. The instruction generation module is used to predict the process evolution trajectory of multiple repair targets within a preset time period based on feature information, transform the optimization trade-off result into the expected co-evolution trajectory and calculate the deviation between the two, and generate control instructions for the MICP grouting equipment based on the deviation. The command control module is used to control the MICP grouting equipment to adjust the grouting parameters according to the control commands.

Images